Skip to content

The model glossary — every FM in the talk, one sentence each

Explainer page for the FM-to-Virtual-Cells talk. The flat quick-reference: every foundation model and virtual cell named anywhere in the talk corpus, one sentence apiece, grouped by family. When a name comes up mid-prep and you just need "what is that again," look here. For the deep version of the 11 anchor models — resources, framework, unique feature, gap exposed — see supplementary §A.

Models marked ★ are the talk's 11 anchor models and link to their full dossier.

Single-cell foundation models & virtual cells

  • scGPT ★ — the sc-FM that defined the category: first to tokenize both genes and cells, enabling cross-cell and cross-gene attention (§A.1).
  • Geneformer V2 ★ — encoder-only BERT with rank-based tokenization; the 104M cancer-curated variant matches the 316M general model at ⅓ the compute, killing the "scale wins" narrative (§A.2).
  • UCE (Universal Cell Embedding) ★ — the cross-species sc-FM: handles 8 species in one embedding space by bridging through ESM2 protein embeddings (§A.3).
  • STATE ★ — Arc Institute's Tahoe-100M-native virtual cell: a State Embedding module plus a State Transition module that predicts perturbation outcomes, the first production virtual cell at 100M-cell scale (§A.4).
  • TranscriptFormer ★ — CZ Biohub's first generative cross-species sc-FM, trained on 112M cells across 12 species spanning 1.53B years of evolution (§A.5).
  • Generative Virtual Cells ★ — a workshop proof-of-concept (<$250 compute) demonstrating the design pattern: a joint planner + world model updated under validation gating, not offline-trained on a frozen snapshot (§A.6).
  • scFoundation — large sc-FM (BioMap) notable for read-depth-aware attention; one of the four models the 2025 reckoning repeatedly tested and found wanting on perturbation prediction.
  • CellPLM — cell-as-token sc-FM (Wen et al., ICLR 2024) that, with scTab, beats scGPT and Geneformer on cell-typing at 10–20× less training cost.
  • scBERT — the early BERT-style single-cell model that pre-dated the 2023 sc-FM wave; mostly of historical interest now.
  • scMulan — a multitask single-cell language model; one of the seven scFMs benchmarked in CellBench-LS.
  • GenePT / scELMo — not trained sc-FMs at all: frozen LLM embeddings of gene names plus a logistic-regression head (~$200 compute) that match or beat scGPT zero-shot on cell-type annotation.
  • xVERSE — the 2026 architectural response: a transcriptomics-native (non-language-model) sc-FM that beats LM-derived sc-FMs by 17.9% on representation and resolves rare cell types from as few as 4 cells.
  • TxPert — Wenkel et al.'s Nature Biotechnology 2026 perturbation-prediction model using multiple knowledge graphs as the inductive bias — "the reckoning answering itself," since Wenkel co-authored the 2025 latent-additive critique.
  • MAP — a knowledge-driven sc-FM offering zero-shot prediction for unprofiled drugs; a 2026 exemplar of biology-specific inductive bias.
  • Nicheformer — Theis lab's spatial-omics sc-FM (~80M params, ~110M dissociated + spatial cells, ~$25k) trained with a niche-aware objective.
  • CellFM — a single-cell foundation model from Zhang et al. (Nature Communications 2025).
  • scPRINT — Kalfon et al.'s sc-FM (Nat Commun 2025) pretrained on 50M cells, aimed specifically at robust zero-shot gene-network inference rather than perturbation prediction.
  • scPRINT-2 — the bioRxiv 2025.12 successor (Kalfon, Peyré & Cantini), pitched as a next-generation cell FM and a benchmark suite for evaluating them.

Pathology foundation models

  • Virchow2 ★ — the current pathology-FM SOTA and the most hardware-transparent: a 632M-param ViT-H/14 trained on ~2B H&E/IHC tiles from ~225k patients (§A.7).
  • UNI2-h ★ — the Mahmood-stack flagship tile encoder (681M params) and the opaque counterpoint to Virchow2; HF-gated since Jan 2025 (§A.8).
  • UNI — the original Mahmood-lab pathology tile encoder (Nature Medicine 2024), predecessor of UNI2-h.
  • Virchow / Virchow2G — the Paige + MSK pathology FM line: Virchow is the V1 tile encoder; Virchow2G is the 1.85B-param scale-up on the same fleet.
  • CHIEF — a pathology FM for cancer detection and prognosis across tissue types.
  • Prov-GigaPath — a whole-slide pathology FM (Microsoft + Providence) trained on real-world clinical slides.
  • Phikon / Phikon-v2 — Owkin's pathology FMs, notable as Apache-2.0 exceptions in a family that is otherwise license-restrictive.
  • H-optimus-0 — Bioptimus's open-weights pathology tile encoder, another permissive-license exception.
  • Hibou — an open pathology FM (Apache-2.0).
  • CONCH — a contrastive vision-language pathology FM from the Mahmood lab.
  • TITAN — the slide-level model in the Mahmood vertical stack, sitting above the UNI2-h tile encoder.
  • PathChat / PathChat-2 / PathChat-DX — the Mahmood lab's vision-language pathology assistant; PathChat-DX is the first generative-AI pathology tool with FDA Breakthrough Designation (Jan 2025).

Genomic foundation models

  • AlphaGenome ★ — DeepMind's variant-effect SOTA: a single U-Net + transformer model that wins 25/26 regulatory variant-effect benchmarks at 1-Mb context (§A.9).
  • Evo2 ★ — Arc + NVIDIA's generative genomic FM (7B/40B params, ~$5M training): the only genomic FM with demonstrated in-context learning and a 1M-token context (§A.10).
  • Nucleotide Transformer — an early DNA foundation model (InstaDeep) pretrained across many genomes for variant and regulatory tasks.
  • DNABERT-2 — a BPE-tokenized DNA FM, the efficiency-focused successor to the original DNABERT.
  • HyenaDNA — a state-space (non-transformer) genomic FM built for very long context at single-nucleotide resolution.
  • Caduceus — a bidirectional, reverse-complement-equivariant state-space DNA FM for long-range genomic tasks.
  • Enformer — the pre-FM-era DeepMind model predicting gene expression from sequence via attention; one of the single-task models AlphaGenome replaces.
  • Borzoi — Calico's RNA-seq-coverage prediction model, an Enformer successor, also subsumed by AlphaGenome's single-model approach.

Protein foundation models

  • ESM-3 ★ — EvolutionaryScale's 98B-param multimodal protein FM with 7 token tracks; generated esmGFP de novo and sets the cleanest compute-disclosure benchmark in biology FMs (§A.11).
  • ESM-2 — the open-weights predecessor protein language model; still the workhorse for most academic reproductions.
  • ESMFold — the ESM-2-based structure predictor: fast single-sequence folding without an MSA.
  • AlphaFold 3 — DeepMind / Isomorphic's structure predictor extended to complexes — proteins with nucleic acids, ions, and ligands.
  • Proteina / Proteina Complexa — generative protein-design FMs; Proteina Complexa reports a 63.5% wet-lab hit rate on a PDGFR binder-design task.

Multimodal / vision-language

  • BioMedCLIP — a biomedical image–text contrastive model trained on large-scale figure–caption pairs.
  • Med-Gemini — Google's medical multimodal model family for clinical reasoning and medical-image Q&A.

Agentic systems built on foundation models

Not foundation models themselves, but they recur throughout the talk because the 2026 story is agentic AI and FMs converging.

  • rBio — a Qwen model post-trained to reason over a virtual cell, using TranscriptFormer as a verifier.
  • VCHarness — an autonomous agent that builds virtual-cell models end to end.
  • CellVoyager — an agent that analyzes single-cell data autonomously (Nature Methods 2026).
  • MedAgentGym — an agentic environment / benchmark for training and evaluating medical AI agents.
  • PathChat-DX — an LLM agent that calls a pathology FM as a tool; the clearest example of "agents use FMs."

Where to go next

Last updated 2026-05-14.